1 / 34

Rate Control For Streaming Video Over Wireless

Rate Control For Streaming Video Over Wireless. Reading: Minghua Chen and Avideh Zakhor , “Rate Control for Streaming Video over Wireless ”, IEEE INFOCOM 2004. 2011. 05. 03 Kim, Dong Joo. Introduction.

donnel
Download Presentation

Rate Control For Streaming Video Over Wireless

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Rate Control For Streaming Video Over Wireless Reading: Minghua Chen and Avideh Zakhor , “Rate Control for Streaming Video over Wireless ”, IEEE INFOCOM 2004. 2011. 05. 03 Kim, Dong Joo

  2. Introduction • Rate control is important to multimedia streaming applications in both wired and wireless networks. • It results in full utilization of bottleneck links by ensuring that sending rates are not too low. • It prevents congestion collapse by ensuring that sending rates are not too aggressive. • Actual network collapse of the Internet happened in the past. • Proper rate control ensures fairness between users sharing common links in a given network.

  3. Introduction • TCP-friendly rate control (TFRC) is equation-based rate control scheme for streaming and is very popular in wired networks • In TFRC the TCP-friendly rate is determined as a function of • packet loss rate • round-trip time (RTT) • packet size • The goal of TFRC is to mimic the long-term steady performance of TCP

  4. Introduction • Three advantages to rate control using TFRC • It can fully utilize bottleneck capacities while preventing congestion collapse. • It is fair to TCP flows, which are the dominant source of traffic on the Internet. • The TFRC results in small rate fluctuation, making it attractive for streaming applications that require constant video quality.

  5. Introduction • The key assumption behind TCP and TFRC is that packet loss is a sign of congestion. • In wireless networks, however, packet loss is dominated by physical channel errors, violating this key assumption. • Neither TFRC nor TCP can distinguish between packet loss due to buffer overflow and that due to physical layer errors. • This results in underutilization of the wireless channel. • Hence, rate control for streaming applications over wireless is still an open problem.

  6. Introduction • There have been a number of efforts to improve the performance of TCP or TFRC over wireless • How to improve the performance? • Hiding end hosts from packet loss caused by wireless channel error. • Providing end hosts the ability to distinguish between packet loss caused by congestion and that caused by wireless channel error.

  7. Introduction • Snoop, a well-known solution, is a TCP aware local retransmission link layer approach • A Snoop module resides on a router or base station on the last hop (i.e., the wireless link) • A Snoop module records a copy of every forwarded packet. • Assuming a Snoop module can access TCP acknowledgment (ACK) packets from the TCP receiver, • it looks into the ACK packets and carries out local retransmissions when a packet is corrupted by wireless channel errors. • While doing the local retransmission, the ACK packet is suppressed and not forwarded to the TCP sender. • These schemes can potentially be extended to TFRC in order to improve performance by using more complicated treatment of the ACK packets from the TFRC receiver.

  8. Introduction • Explicit loss notification (ELN) can also be applied to notify a TCP/TFRC sender when packet loss is caused by wireless channel errors rather than congestion • TFRC can take into account only the packet loss caused by congestion when adjusting the streaming rate. • End-to-end statistics can be used to help detect congestion when a packet is lost • For example, by examining trends in the one way delay variation, Parsa and Garcia-Luna Aceves interpret loss as a sign of congestion if one-way delays or queuing delay are increasing, and a sign of wireless channel error otherwise. • Similarly, Barman and Matta proposed a loss differentiation scheme based on the assumption that the variance of RTT is high when congestion occurs, and low otherwise.

  9. Introduction • Cen et al. present an end-to-end-based approach to facilitate streaming over wireless • They combine packet inter-arrival times and relative one-way delay to differentiate between packet loss caused by congestion and that due to wireless channel errors. • There are two key observations behind their approach • relative one way delay increases monotonically if there is congestion • inter-arrival time is expected to increase if there is packet loss caused by wireless channel error. • These two statistics can help differentiate between congestion and wireless errors. • However, the high wireless error misclassification rate may result in underutilizing the wireless bandwidth • There are other people who proposed a similar approach to improve video streaming performance in the presence of wireless error, under the assumption that the wireless link is the bottleneck.

  10. Introduction • Other schemes that use end-to-end statistics to detect congestion, can also be combined with TFRC for rate control. • The congestion detection scheme can be used to determine whether or not an observed packet loss is caused by congestion; TFRC can then take into account only those packet losses caused by congestion when adjusting the streaming rate. • Tang et al. proposed an idea of using small dummy packets to actively probe whether the network is congested in case of packet loss. • Yang et al. proposed a cross-layer scheme that uses link layer information to determine whether a packet loss is caused by channel error or congestion, assuming that only the last link is wireless. • In this approach, when a packet is lost, TFRC goes beyond layering abstraction and queries the link layer about recent signal strength. • The packet loss is recognized as due to wireless channel error if recent signal strength is low, and due to congestion otherwise.

  11. Introduction • The disadvantage of end-to-end statistics-based approaches is that • congestion detection schemes based on statistics are not sufficiently accurate. • congestion detection schemes require either cross layer information or modifications to the transport protocol stack. • Another alternative is to use non-loss based rate control schemes. • For instance, TCP Vegas, in its congestion avoidance stage, uses queuing delay as a measure of congestion. • could be designed not to be sensitive to any kind of packet loss, including that due to wireless channel error. • It is also possible to enable routers with ECN marking capability to do rate control using ECN as the measure of congestion • As packet loss no longer corresponds to congestion, ECN-based rate control does not adjust sending rate upon observing a packet loss.

  12. Introduction • In this article • The necessary and sufficient condition under which using one TFRC connection in wireless streaming applications results in underutilization of the wireless bandwidth are explored. • The use of multiple simultaneous TFRC connections for a given wireless streaming application is proposed. • The advantages of proposed approach are as follows • It is an end-to-end approach, and does not require any modifications to network infrastructure and protocols, except at the application layer. • It has the potential to fully utilize the wireless bandwidth, provided the number of connections and packet size are selected appropriately.

  13. Problem Formulation • Setup And Assumptions • S in the wired network is streaming video to r in the wireless network. • The wireless link is assumed to have available bandwidth Bw, and packet loss rate Pw • There could also be packet loss caused by congestion at node 2, denoted by Pc • The end-to-end packet loss rate observed by receiver is denoted by P • The streaming rate is denoted by T • Wireless channel is underutilized if the streaming throughput is less than the maximum possible throughput over the wireless link • T(1-p) < Bw(1-pw) s : video server in the wired network r : receiver Bw : available bandwidth Pw : packet loss rate caused by wireless channel error

  14. Problem Formulation • Given this scenario, the following are assumed • There is no cross traffic at either node 1 or node 2 • In the long term, the wireless link is assumed to be the bottle neck. • by this there is no congestion at node 1 • There is no congestion and queuing delay at node 2 if and only if wireless bandwidth is underutilized (i.e., pc = 0 and minimum RTT, defined as RTTmin, if and only if T < Bw ) • Bwand pware assumed to be constant, at least on the timescale in which the analysis is carried; packet loss caused by wireless channel error is assumed to be random and stationary • For simplicity, the backward route is assumed to be error-free and congestion-free • Based on this scenario, two goals can be stated as follows • The streaming rate should not cause any network instability (i.e., congestion collapse) – TFRC can clearly meet • It should lead to the optimal performance (i.e., highest possible throughput and lowest possible packet loss rate)

  15. Problem Formulation • A Sufficient and Necessary Condition For Underutilization • The following model is used for TFRC in the analysis • The overall packet loss rate is p, a combination of pw and pc, and can be written as • This shows that pwis a lower bound for p, and that the bound is reached if and only if there is no congestion (i.e., pc= 0). • Combining this observation and Eq. 1, and upper bound, Tb, on the streaming rate of one TFRC connection can be derived as follows: T : the sending rate S : the packet size rtt : the end-to-end RTT P : the end-to-end packet loss rate k : a constant factor

  16. Problem Formulation • If there is no congestion (i.e., pc= 0), and hence no queuing delay caused by congestion, we get rtt = RTTmin, p= pw, and Tb achieves the upper bound T = Tbin Eq. 3. • In this case the throughput is Tb(1 – pw), which is the upper bound of throughput given one TFRC connection for the scenario shown in Fig. 1. • Based on these, it is possible to state the following: • Theorem 1: Given the scenario and assumptions earlier, a sufficient and necessary condition for one TFRC connection to underutilize a wireless link is • This implies that if the available bandwidth is larger than the highest sending rate one TFRC can achieve, underutilization happens. • Eq. 4. is not satisfied through modifications to network infrastructure or protocols.

  17. Problem Formulation • A Strategy To Reach Optimal Performance • It is not necessary to avoid the condition in Eq. 4 in order to achieve reasonable performance for one application because it is conceivable to use multiple simultaneous connections for one application. • The total throughput of the application is expected to increase with the number of connections until it reaches the hard limit of Bw(1 – pw). • multiple connections can be used to achieve optimal performance: throughput of Bw(1 – pw) and packet loss rate of pw. • By opening one TFRC connection with packet size S, the application achieves a throughput of • and packet loss rate of pw. This is because according to Theorem 1, underutilization implies rtt = RTTmin, p = pw, and

  18. Problem Formulation • The case with two TFRC connections from sender s to receiver r in Fig.1. • since pw for each of the two TFRC connections remain unchanged from the case with one TFRC connection, the throughput upper bound for each of the two TFRC connections is • and the aggregate throughput upper bound for both of them is • which is smaller than Bw(1 – pw), implying channel underutilization and no congestion. Consequently, end-to-end packet loss rate p is at pw, and the total throughput for both connections is

  19. Problem Formulation • A similar argument can be repeated with three TFRC connections, except that the wireless channel is no longer underutilized and rtt >RTTmin. • Furthermore, if the buffer on node 2 overflows, pcwill no longer be zero; hence, using Eq. 2 we get p > pw • In this case, the wireless link is still fully utilized at T(1-p) = Bw(1- pw), but RTT is no longer at the minimum value RTTmin, and overall packet loss rate p could exceed pw (i.e. the overall packet loss rate in the two connections case) • In general, given Bw, pw, and packet size S for each connection, it can be shown that when full wireless channel utilization occurs, the optimal number of connections, nopt, satisfies

  20. Problem Formulation • Thus, what really matters is the product of noptand S; as such, it is always possible to achieve full wireless channel utilization by choosing noptto be an integer, and selecting S accordingly. • It is also possible to analyze the case with different packet sizes for different connections, but it is not fundamentally different from the case with the same packet size for all connections. So, S is fixed in this article. Then, the optimal number of connections is given byresulting in throughput of and packet loss rate of pw. Opening more than nopt connections results in larger rtt, or possibly higher end-to-end packet loss rate.

  21. Problem Formulation • To summarize • If the number of TFRC connections is too small, wireless channel becomes under-utilized. • If the number of connections is chosen optimally based on Eq.5, then wireless channel becomes fully utilized, the total throughput becomes Bw(1 – pw), with rtt = RTTmin, and the overall packet loss rate achieves the lower bound pw • If the number of connections exceeds nopt, even though the wireless channel continues to be fully utilized, the rtt will increase beyond RTTminand later on packet loss rate can exceed the lower bound pw. • as number of connections exceeds nopt, the sending rate of each connection has to decrease. Thus by Eq.1, the product has to increase, so either rtt increases or p increases, or they both increase.

  22. Problem Formulation • a strategy leading to optimal performance. • keep increasing the number of connections until an additional connection results in an increase of end-to-end round trip time or packet loss rate.

  23. The Proposed Solution: Multiple TFRC • The basic idea behind MULTFRC is • to measure the round trip time • to adjust the number of connections accordingly • The purpose of MULTFRC is • to utilize the wireless bandwidth efficiently • to ensure fairness between applications. • Two components in the proposed system • an rtt measurement subsystem(RMS) • a connections controller subsystem(CCS) • both of them reside at the sender

  24. The Proposed Solution: Multiple TFRC • an rtt Measurement Subsystem(RMS) • measures average rttover a window, denoted ave_rtt, and reports it to the CCS. • specifically, RMS receives average rttsample , measured in the past RTT window, from the receiver every RTT. • then further computes a smoothed version of these average rttsevery m reports (i.e., ). • one can set m to be large values to reduce the noise in ave_rtt, or small values to make the system more responsive to changes in RTT.

  25. The Proposed Solution: Multiple TFRC • a Connections Controller Subsystem(CCS) • Inspired by TCP, CCS’s basic functionality is to inversely increase and additively decrease(IIAD( , )) the number of connections n, based on the input from RMS with and being preset constant parameters. • Specifically, it first sets the rtt_min as the minimum ave_rtt seen so far, and then adapts the number of connection n as follows: • The reason for this is fair and efficient sharing among multiple MULTFRC applications, and between MULTFRC and TCP or TFRC connections

  26. The Proposed Solution: Multiple TFRC • How to control the number of connections? • For a given route, ave_rtt – rtt_min corresponds to current queuing delay, and rtt_min is a threshold on the queuing delay that MULTFRC can tolerate before it starts to decrease the number of connections. • Ideally, ave_rtt becomes larger than rtt_min if and only if the link is fully utilized, and the queue on bottleneck link router is built up, introducing additional queuing delay. • Thus by evaluating the relation between ave_rtt and rtt_min, MULTFRC detects full utilization the wireless link, and controls the number of connections accordingly.

  27. The Proposed Solution: Multiple TFRC • Additionally • When there is a route change due to either change in the wireless base station or route change within the wired Internet, the value of rtt_min changes, affecting the performance of MULTFRC. • Under these conditions, it is conceivable to use route change detection tools such as ‘traceroute’ to detect the route change, in order to reset rtt_min to a new value. • Since MULTFRC always keeps at least one connection open, the overall throughput of MULTFRC will not go to zero.

  28. The Proposed Solution: Multiple TFRC • Performance evaluation by NS-2 simulations & actual experiments • Actual experiments were done over Verizon Wireless 1xRTT CDMA data network. • Simulations showed that MULTFRC can achieve reasonable utilization of the wireless bandwidth, and does not starve applications that use one TCP connection. • For actual experiments over lxRTT, the followings were set up • A desktop server connected to the Internet via 100 Mb/s Ethernet • A laptop receiver connected to the Internet via the Verizon Wireless lxRTT CDMA data network(It is the bottleneck for the streaming connection). • The lxRTT CDMA data network is advertised to operate at data speeds of up to 144 kb/s for one user. (actually 80 ~ 97 kb/s confirmed by experiment) • The packet size S is 1460 bytes. • As pw cannot be controlled, the average throughput, average number of connections, and packet loss rate were measured.

  29. The Proposed Solution: Multiple TFRC • Comparison the performance of the MULTFRC system and one TFRC connection • MULTFRC on average opens 1.8 connections, and results in 60 percent higher throughput at the expense of a larger RTT and higher packet loss rate. • packet loss details of MULTFRC for a 30-min-long experiment with packet size of 760 bytes. • Both the packet loss rate and burstiness of the loss increase as the number of connections increases.

  30. Video Streaming Simulations • Performance evaluation of MULTFRC in video streaming application • Simulated streaming of a 60 s video clip through a channel, with throughput trace corresponding to one of the traces obtained from actual experiments over lxRTT CDMA as described earlier. • Goal is to compare the quality of video streaming achievable using one TFRC connection with that of MULTFRC. • For experiment, the followings were set up • Encoded 300 frames of the news.cif sequence using MPEG4 at bit rates varying from 50 to 100 kb/s3 • The frame rate is 10 frames/s • the intra-frame refresh rate is once every 15 frames. • The coded video bit stream is packetized with fixed packet size of 760 bytes. • The packets are then protected using Reed-Solomon (RS) codes with different protection levels for one TFRC and MULTFRC. • The RS-coded packets are then passed through channels simulated using one TFRC, and MULTFRC packet level traces each lasting 70 seconds, selected from the 30 min actual experiments

  31. Video Streaming Simulations • The throughput and packet loss details for a 70 second segment of one TFRC and MULTFRC connections are shown in following figure. • both throughput and packet loss rate are higher for MULTFRC than for one TFRC case. • The large throughput fluctuations in MULTFRC due to changing number of connections can potentially be argued not to be suitable for video applications in general; however, proper buffering can absorb these fluctuations in non-delay-sensitive streaming applications.

  32. Video Streaming Simulations • To show the efficiency of MILTFRC, the playback buffer occupancies of MULTFRC and one TFRC for several bit rates were compared. • compared toone TFRC case, MULTFRC can sustain video streaming at higher average bit rates and hence higher visual quality, despite the fact that it needs stronger forward error correction (FEC) to combat the higher packet loss rate.

  33. Discussion And Future Work • Other work similar but unrelated to the proposed approach includes MULTCP and NetAnts which open multiple connections to increase throughput. • MULTCP was originally proposed to provide differential service, and was later proposed to improve the performance in high bandwidth-RTT product network. • NetAnts achieves higher throughput by opening multiple connections to compete for bandwidth against others’ applications. • The differences between NetAnts and our approach are as follows. • opening more connections than needed in wired networks unnecessarily increases the end-to-end packet loss rate experienced by an end host • unlike the proposed approach, there is no mechanism to control the number of connections in NetAnts. • Future work includes the stability, scalability, and fairness analysis of the proposed approach.

  34. Thank you Kim, Dong Joo Hankuk University of Foreign Studies Dept. of Information and Communications Engineering sangnamleader@hufs.ac.kr

More Related